Efficient encoder-decoder model for cell segmentation in 3D microscopy images. 3D microscopy images are analyzed slice-wise as stacks of 2D slices. Context from adjacent 2D slices is encoded using our pseudocoloring algorithm (see below).
Results in the Cell Segmentation Benchmark:
| Dataset | Team / Method | Model(s) | #params ⇓ | SEG Score ⇑ | Ranking* |
|---|---|---|---|---|---|
| Fluo-C3DH-A549-SIM | DKFZ-GE [Repo] | 4x 3D nnU-Net | 176.0M | 0.955 | 1/12 |
| Ours | EfficientCellSeg | 6.7M | 0.951 | 2/12 | |
| Fluo-N3DL-TRIC | KIT-Sch-GE [Repo] | Dual U-Net | 46.4M | 0.821 | 1/8 |
| Ours | EfficientCellSeg | 6.7M | 0.782 | 3/8 | |
| Fluo-C3DL-MDA231 | KIT-Sch-GE [Repo] | Dual U-Net | 46.4M | 0.710 | 1/19 |
| Ours | 2x EfficientCellSeg | 13.4M | 0.646 | 2/19 |
Context from adjacent z-slices - an approximation of regions where cells might be located in these slices - is determined by CLAHE filtering and thresholding. Afterwards, these regions are highlighted in the current z-slice via a multiply-accumulate operation. Context from the previous z-slice (z - 1) is highlighted in the red channel, context from the next z-slice (z + 1) in the blue channel. The result is a pseudocolor image that is similar to natural color images in the sense that it still shows the same scene in all three color channels with moderate differences between the channels. We assume that this similarity is the cause of the good performance in combination with ImageNet weights.
EfficientCellSeg: Efficient Volumetric Cell Segmentation Using Context Aware Pseudocoloring, Wagner, Royden and Rohr, Karl, MIDL 2022; arXiv (arXiv:2204.03014)
@inproceedings{wagner2022efficientcellseg,
title={EfficientCellSeg: Efficient Volumetric Cell Segmentation Using Context Aware Pseudocoloring},
author={Royden Wagner and Karl Rohr},
booktitle={Medical Imaging with Deep Learning},
year={2022}
}